PreVA: Predictive Vertical Autoscaler Using Multi Bi-GRU for Sustainable Cloud-Native Computing

초록

Container resource autoscaling techniques efficiently manage container resources configured in a cloud-native computing environment. The vertical autoscaling (VA) technique provides resource elasticity by resizing a container resource in response to a generated load. However, VA techniques demonstrate inefficient scaling performance for workloads with patterns that differ from past ones because they only consider patterns of past resource usage and operate based on reactive mechanisms. Additionally, the service is temporarily disrupted by deleting and recreating containers when resources are resized. Therefore, this study proposes a predictive vertical autoscaler (PreVA) that efficiently utilizes resources and ensures service sustainability under various workload patterns. PreVA extracts temporal features from collected CPU and memory usage metrics and then trains a multi bidirectional gated recurrent unit model to forecast future resource usage with high accuracy. PreVA also utilizes forecasted resource usage to calculate optimal resource sizes for future workloads. Finally, PreVA performs rolling updates to resize resources and ensure service sustainability. PreVA is validated by performing offline simulations in a cloud-native computing environment, with approximately 90% resource utilization for various workloads. Additionally, compared with existing VA techniques, PreVA reduces the number of resource overloads and service disruptions by up to 40 and 409, respectively.

키워드